pyterrier_adaptive
PyTerrier implementation of Adaptive Re-Ranking using a Corpus Graph (CIKM 2022).
Getting Started
Install with pip:
pip install --upgrade git+https://github.com/terrierteam/pyterrier_adaptive.git
Basic Example over the MS MARCO passage corpus (making use of the pyterrier_t5 and pyterrier_pisa plugins):
import pyterrier as pt
from pyterrier_t5 import MonoT5ReRanker
from pyterrier_pisa import PisaIndex
from pyterrier_adaptive import GAR, NpTopKCorpusGraph
dataset = pt.get_dataset('irds:msmarco-passage')
retriever = PisaIndex.from_dataset('msmarco_passage').bm25()
scorer = pt.text.get_text(dataset, 'text') >> MonoT5ReRanker(verbose=False, batch_size=16)
graph = NpTopKCorpusGraph.from_dataset('msmarco_passage', 'corpusgraph_bm25_k16').to_limit_k(8)
pipeline = retriever >> GAR(scorer, graph) >> pt.text.get_text(dataset, 'text')
pipeline.search('clustering hypothesis information retrieval')
# qid query docno rank score iteration text
# 1 clustering hypothesis information retrieval 2180710 0 -0.017059 0 Cluster analysis or clustering is the task of ...
# 1 clustering hypothesis information retrieval 8430269 1 -0.166563 1 Clustering is the grouping of a particular set...
# 1 clustering hypothesis information retrieval 1091429 2 -0.208345 1 Clustering is a fundamental data analysis meth...
# 1 clustering hypothesis information retrieval 2180711 3 -0.341018 5 Cluster analysis or clustering is the task of ...
# 1 clustering hypothesis information retrieval 6031959 4 -0.367014 5 Cluster analysis or clustering is the task of ...
# .. ... ... ... ... ... ...
# iteration column indicates which GAR batch the document was scored in ^
# even=initial retrieval odd=corpus graph -1=backfilled
Evaluation on a test collection (TREC DL19):
from pyterrier.measures import *
dataset = pt.get_dataset('irds:msmarco-passage/trec-dl-2019/judged')
pt.Experiment(
[retriever, retriever >> scorer, retriever >> GAR(scorer, graph)],
dataset.get_topics(),
dataset.get_qrels(),
[nDCG, MAP(rel=2), R(rel=2)@1000],
names=['bm25', 'bm25 >> monot5', 'bm25 >> GAR(monot5)']
)
# name nDCG AP(rel=2) R(rel=2)@1000
# bm25 0.602325 0.303099 0.755495
# bm25 >> monot5 0.696293 0.481259 0.755495
# bm25 >> GAR(monot5) 0.724501 0.489978 0.825952
Reproduction
Detailed instructions to come!
Building a Corpus Graph
You can construct a $k$ corpus graph using any retriever transformer and a corpus iterator.
Example:
from pyterrier_adaptive import CorpusGraph
from pyterrier_pisa import PisaIndex
dataset = pt.get_dataset('irds:msmarco-passage')
# Build the index needed for BM25 retrieval (if it doesn't already exist)
idx = PisaIndex('msmarco-passage.pisa', threads=45) # adjust for your resources
if not idx.built():
idx.index(dataset.get_corpus_iter())
# Build the corpus graph
K = 16 # number of nearest neighbours
graph16 = CorpusGraph.from_retriever(
idx.bm25(num_results=K+1), # K+1 needed because retriever will return original document
dataset.get_corpus_iter(),
'msmarco-passage.gbm25.16',
k=K)
You can load a corpus graph using the .load(path) function. You can simulate lower $k$ values
using .to_limit_k(k)
graph16 = CorpusGraph.load('msmarco-passage.gbm25.16')
graph8 = graph16.to_limit_k(8)
Citation
Adaptive Re-Ranking with a Corpus Graph. Sean MacAvaney, Nicola Tonellotto and Craig Macdonald. In Proceedings of CIKM 2022.
@inproceedings{gar2022,
title = {Adaptive Re-Ranking with a Corpus Graph},
booktitle = {Proceedings of ACM CIKM},
author = {Sean MacAvaney and Nicola Tonellotto and Craig Macdonald},
year = 2022
}
Metadata
Release files for pyterrier-adaptive 0.2.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| pyterrier_adaptive-0.2.1.tar.gz | 14.4 kB | Without provenance | Details for pyterrier_adaptive-0.2.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| pyterrier_adaptive-0.2.1-py3-none-any.whl | Python 3 | none | any | Details for pyterrier_adaptive-0.2.1-py3-none-any.whl |
Total release size: 28.1 kB
Release files / pyterrier_adaptive-0.2.1.tar.gz
| Download URL | pyterrier_adaptive-0.2.1.tar.gz |
|---|---|
| Size | 14.4 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
f3a34af73d09d6c47e94336803c428c90afeee3dbafa7182ec6a3d111e3a0c5c
|
|
BLAKE2b-256 checksum How to use checksums |
e947f2e076cf5d176079eb88446c1a84594445ba9c166c7416ba6473ecd3b0e1
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.14.5
|
Release files / pyterrier_adaptive-0.2.1-py3-none-any.whl
| Download URL | pyterrier_adaptive-0.2.1-py3-none-any.whl |
|---|---|
| Size | 13.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
1c087db3d9fb830b158f60f017a89a04b43aad66467765c3439a99d9c3b37152
|
|
BLAKE2b-256 checksum How to use checksums |
034893ceb8988876481ae9523a0c22b07564f43de0537e9368e9ff22728c2f9c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.14.5
|
Security for pyterrier-adaptive 0.2.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.